Verifying reward model scores for trustworthy large language model alignment
VRMaaS: Verifiable Reward-Model-as-a-Service for Trustworthy LLM Alignment
Cryptography and Security
Summary
Large language models (LLMs) use reward models to measure how well they perform certain tasks. But verifying that these reward scores really come from the correct model and settings can be tricky. The authors propose a system that attaches a proof to each reward score, allowing auditors to check its authenticity without needing to rerun the entire model. This method works efficiently with different model sizes and can be used in audits or governance where trust in the model's output is critical.
What this means in practice
- •For software developers: Verify reward scores from LLM alignment pipelines without rerunning private models, enabling more trustworthy AI services.
- •For blockchain engineers: Use zero-knowledge proofs to partially attest to model reward outputs on-chain to support decentralized governance or payment systems.
Authors
Guofu Liao, Hailang Xi, Taotao Wang, Shengli Zhang, Rongcheng Tu, Hao Liao
Abstract
Reward models (RMs) are increasingly deployed as service components in LLM alignment pipelines, where payment, audit, moderation, and governance workflows act on the scores they return. An inference proof alone does not show whether a score belongs to the intended request, model checkpoint, numerical configuration, and verifier: in a controlled comparison that isolates the accepted statement from prover scale, a generic inference proof wrapped for on-chain checking remains valid after six classes of such service-field substitution. We formulate verifiable reward-model-as-a-service (VRMaaS) as a service-output integrity problem and present a framework in which each returned reward becomes a proof-carrying artifact. Its off-chain acceptance rule binds request identity, committed weights, verifier and configuration identifiers, proof-task order, and the digest of the exported package, so auditors can check a reward without rerunning the private RM or exposing inputs, checkpoints, or hidden activations; an optional on-chain zero-knowledge wrapper gives contracts a partial attestation of selected consistency checks and supplied service fields. The framework decomposes Transformer RM serving into arithmetic and lookup-backed proof tasks, aggregates them into a composite proof, and links hidden operands to the committed weights. A GPU prototype on 1B, 3B, and 8B reward models over 128-1024 tokens preserves 93.5-99% pairwise agreement with the baseline RM and rejects every tested service-field substitution in complete artifacts. Using exact integer witnesses and request-level aggregation of weight-binding openings, it generates complete artifacts in 49.8-206.1 seconds across the evaluated settings; auditors verify retained artifacts in 1.94-5.11 seconds, and partial public attestation costs about 217K gas. These costs suit delayed audit, settlement, and governance workflows.